less than 1 minute read

Reference

SuperPoint: Self-Supervised Interest Point Detection and Description

TILDE: A Temporally Invariant Learned DEtector

The TILDE interest point detection system used a principle similar to Homographic Adaptation; however, their approach does not benefit from the power of large fully-convolutional neural networks.

Homographic Adaptation ์ด๋ž€?

Homographic Adaptation์€ SuperPoint์—์„œ ์ œ์•ˆ๋œ ๊ธฐ๋ฒ•์œผ๋กœ, ๊ด€์‹ฌ์  ๊ฒ€์ถœ์˜ ๋ฐ˜๋ณต์„ฑ์„ ๋†’์ด๊ณ  ๋„๋ฉ”์ธ ๊ฐ„ ์ ์‘(์˜ˆ: ํ•ฉ์„ฑ์—์„œ ์‹ค์ œ ์ด๋ฏธ์ง€๋กœ)์„ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์œ„ํ•œ ๋‹ค์ค‘ ์Šค์ผ€์ผ, ๋‹ค์ค‘ ํ˜ธ๋ชจ๊ทธ๋ž˜ํ”ผ ์ ‘๊ทผ ๋ฐฉ์‹์ž…๋‹ˆ๋‹ค.

์ฃผ์š” ํŠน์ง•์€ ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค:

  • ๋‹ค์ค‘ ๋ณ€ํ™˜: ์ž…๋ ฅ ์ด๋ฏธ์ง€์— ์—ฌ๋Ÿฌ ๋ฌด์ž‘์œ„ ํ˜ธ๋ชจ๊ทธ๋ž˜ํ”ผ๋ฅผ ์ ์šฉํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์‹œ์ ๊ณผ ์Šค์ผ€์ผ์—์„œ ์žฅ๋ฉด์„ ๋ณผ ์ˆ˜ ์žˆ๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค.
  • ์ž๊ธฐ ์ง€๋„ ํ•™์Šต: ๋ ˆ์ด๋ธ”์ด ์—†๋Š” ๋Œ€์ƒ ๋„๋ฉ”์ธ์˜ ์ด๋ฏธ์ง€์— ์ž๋™์œผ๋กœ ๋ ˆ์ด๋ธ”์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค.
  • ์„ฑ๋Šฅ ํ–ฅ์ƒ: ์ดˆ๊ธฐ MagicPoint ๊ฒ€์ถœ๊ธฐ์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œ์ผœ ๋” ํ’๋ถ€ํ•œ ๊ด€์‹ฌ์  ์ง‘ํ•ฉ์„ ๋ฐ˜๋ณต์ ์œผ๋กœ ๊ฒ€์ถœํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค.
  • Pseudo ground truth ์ƒ์„ฑ: ์ด ๊ณผ์ •์„ ํ†ตํ•ด ์‹ค์ œ ์ด๋ฏธ์ง€์— ๋Œ€ํ•œ pseudo ground truth ๊ด€์‹ฌ์ ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค.
  • ๋„๋ฉ”์ธ ์ ์‘: ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ์—์„œ ์‹ค์ œ ์ด๋ฏธ์ง€๋กœ์˜ ๋„๋ฉ”์ธ ์ ์‘์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค.

Homographic Adaptation์€ SuperPoint ๋ชจ๋ธ์˜ ํ•ต์‹ฌ ๊ตฌ์„ฑ ์š”์†Œ๋กœ, ๋‹ค์–‘ํ•œ ํ™˜๊ฒฝ์—์„œ์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ณ  ์‹ค์ œ ์„ธ๊ณ„ ์‘์šฉ์— ๋” ์ ํ•ฉํ•œ ๊ด€์‹ฌ์  ๊ฒ€์ถœ๊ธฐ๋ฅผ ๋งŒ๋“œ๋Š” ๋ฐ ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•ฉ๋‹ˆ๋‹ค.

Leave a comment